Intel Corp has made it widely known that it will continue its enormously successful iAPX-86 microprocessor line by adding upwardly-compatible, ever more powerful successors. We are currently on the 80486, which in due course will be followed by an 80586. But what then? Ted Hoff got the whole thing started back in 1971 when he presented Intel with the idea that led to the introduction of the four chip Intel 4004 set. It had a 4-bit data bus, could address 4.5Kb of memory, and had 45 instructions. Two decadec on, and the same company is talking in terms of a microprocessor that has upwards of 2m times the MIPS rating – although the underlying rationale is essentially the same. The 80486 is very much Intel’s sacred cash cow at the moment. Although rival chip-maker Motorola Inc tried to steal a march on Intel by announcing its 68040 processor – the same generation as the 80486 – just a few days before Intel lifted the veil on its hottest new part back in April last year, the company has been unable to compete with Intel’s lead times to silicon. Apricot Computers had an 80486 system out later in the same year but problems with the 68040 have sent shipment dates slipping and sliding, leaving system builders struggling along with only sample quantities of the chip.
Volume deliveries of the 68040, even to Motorola’s largest customers like Hewlett-Packard will not begin before this autumn. Meanwhile Intel has been quietly putting the finishing touches to its plans for a microprocessor architecture dubbed Micro 2000 that will take it up to the turn of the century and beyond. By this time the company envisages a part with up to 100m transistors on a 1 die with 2Mb of cache, running at 250MHz, performing 2,000 MIPS – and fully compatible with the 80386. Going by the release timetable for the iAPX-86 line, this is around the time an 80786 would be due to make an appearance. The current 80486 has 1.2m transistors, and the 80786 will be preceded by an 80586 chip in 1992 with 4m or 5m transistors, and an 80686 in 1996 with many more transistors again. Intel is confident that a steady and predictable stream of technological advances will deliver practical devices for achieving its goals. The total number of transistors that can be crammed onto a given area has been steadily increasing, resulting in a continuous reduction in device dimensions. Intel says it has already produced sub-micron components in the labs – down to 0.1 micron size. It is also relying on improved manufacturing processes to reduce the number of defects per square centimetre of silicon, making larger die sizes possible, and with larger parts, an increase in the number of transistors per die to around 50m by the year 2000, and 100m in the future. Its vision is a four-processor CPU, each having 5m transistors, two vector units with another 5m transistors apiece, another 5m used to make up a self-test unit, 2m for the bus unit, 40m for cache requirements and a further 20m for delivering the user interface, including Intel’s Digital Video Interface technology.
By William Fellows
By comparison, even though one of today’s 80386 personal computers has only about 525,000 transistors, its power is equivalent to that of a minicomputer that would have cost up to $250,000 five or six years ago. The possible clock rate is dependent on a range of factors: power consumption, packaging, interconnect technology and transistor current drive. The faster the thing goes and the more power it consumes, the hotter it gets and the more unreliable it becomes. Intel says it will be using four or more multiple layer interconnections to optimise clock signal routing within the chip, on material with a lower restivity than those used today. With smaller component sizes, increased drive capability and channel lengths as small as 0.1 micron, the firm expects to realise clock rates in excess of 250MHz. Intel stands by a maxim that one of its researchers Pat Gelsinger – outlined in the Micro 2000 white paper – every good idea proven in the mainframe or minicomputer migrates onto the
microprocessor – citing memory management on mainframes in the 1970s that was delivered on the 80286, and the paged virtual memory on late 1970s VAXes that is now found in most commercial microprocessors. The integration of these features into microprocessors occurs – or is focused – when successive performance bottlenecks are found. With each generation of mainframes, the next system bottleneck migrates onto the microprocessor. Intel predicts the next refugee will be multiprocessors. Multiple processors are used on mainframes and minis – now even on micros – to extend performance, eventually multiple CPUs will be incorporated on to a single die. Furthermore, it is already possible to execute two or three instructions per clock cycle today, and Intel is working on getting down to 0.2 or 0.3 clocks per instruction by the turn of the century. These architectural enhancements, together with a 250MHz clock cycle should mean a single CPU will perform at around 700 MIPS – four on a single die going to 2,000 MIPS. If all this makes existing microprocessors look like museum pieces – before we’ve even got properly to grips with them – what is to become of the vast quantities of application software that are currently available for the iAPX-86 chip family? Do we throw it all away, or re-write the whole lot when Micro 2000, 80786 or whatever it will be called eventually arrives? Will Intel be able to offer compatibility across this extended time span?
East German Trabant
For an industry that loves to draw on nomenclature from the motor world, it’s like comparing a smoky old East German Trabant with the newest and fastest model BMW can offer. But even they are now driving side by side on the same autobahns – despite the fact that the two Germanys are aeons apart in technological development. Intel maintains that the burden of compatibility will not be a problem. It reckons that the amount of on-chip space that’ll become available with the advances described should mean that compatibility with previous implementations will be a problem that shrinks in size as far a processor development is concerned. In addition it should allow what are now off-chip features – such as local area network and disk controllers, serial and parallel ports – to be integrated on a single processor, which would see all of a personal computer’s present functionality implemented on a single chip – that could be done now with the 8086 if not the 80286: why hasn’t Intel done it already? The massive increases in power that these developments look set to deliver will, says Intel’s microcomputer components group president Dave House, give a tremendous boost to the scope for developing – and using artificial intelligence, neural networks and, most importantly, the friendly user interfaces that have been missing for so long.